Hierarchical Time Series Forecasting of COVID-19 Cases Using County-Level Clustering Data

In light of the far-reaching impact of the COVID-19 pandemic, the accurate estimation of infected cases, fatalities, and recoveries has become crucial. While much research has focused on national-level predictions, it has become evident that an exclusive focus on broader statistics may lead to inadequate preparedness in managing hospital resources in rural and smaller regions. Given the critical role local areas play in the spread of COVID-19 and the hierarchical structure of available data, this study proposes a novel modeling framework using hierarchical time series forecasting (HTSF) tailored to county-level clusters within the USA. This approach aims to provide short-term daily forecasts for every county in the USA, employing bottom-up, top-down, and minimum trace optimal reconciliation techniques. A major barrier to accurate short-term forecasting is the scarcity of COVID-19 hospitalization data in both rural and urban regions across the USA. To address this limitation, the study employs county-level clustering, a method that enables the generation of forecasts even for areas with limited or no available data. The primary aim is to improve the accuracy of COVID-19 case forecasts by combining autoregressive integrated moving average (ARIMA) models as base forecasts with the HTSF approach. The findings reveal that the bottom-up HTSF method offers comparable performance at the county level but significantly outperforms other approaches at the cluster and national levels for 3-week-ahead forecasting. This highlights the vital role of local regions in achieving more precise and effective pandemic prediction strategies.